Blog - ArchivesPosts of 09/09/2026
An AI that "talks cellular"
- by
cronywell
|
🧬 🤖 SCIENCE AND TECHNOLOGY An AI that "talks cellular" This is IRIS, the model that deciphers the code of signals with which cells decide which tissue to become Fuente: National Geographic España · Estudio original: Nature Methods (Whitehead Institute / MIT) ⏱️ Estimated reading time: 8 minutes 📅 September 2026 |
🔬 AI & HEALTH
An AI that "talks cellular": this is how it predicts which tissue each cell will become
A team at the MIT-affiliated Whitehead Institute has developed IRIS, an artificial intelligence model capable of decoding the chemical language with which cells communicate during embryonic development. By analyzing a cell's gene activity, the system can reconstruct what signals it received and predict exactly which specific tissue — heart, lung, gut, muscle, or spinal cord — it will eventually become, according to National Geographic. The finding, published in Nature Methods, opens the door to large-scale mapping of how a body is built, cell by cell.
🖼️ View image: Human embryonic stem cells — Wikimedia Commons (Public domain/CC, Wikimedia Commons)
🗣️ The secret language of cells
When an embryo begins to develop from a small group of stem cells, these cells, which are in principle identical and have no defined function, begin to specialize: some will become neurons, others liver cells, others muscle fibers. That decision depends not on an isolated internal plan, but on an ongoing conversation: the cells send and receive chemical signals to and from their neighbors, which tell them where they are in the embryo, what stage of development they are in, and what they should become.
These messages travel through multi-step sequences called signaling pathways, authentic biochemical circuits that translate an external stimulus into very specific changes in genetic activity within the cell. Reconstructing this sequence of instructions – what signals a cell was exposed to and at what time – would allow us to understand in depth how tissues are formed and why, when this process fails, diseases appear.
The problem is that, until now, this reconstruction seemed almost impossible to scale. Science assumed that the effects of each signaling pathway varied greatly from one cell type to another, which forced each pathway to be mapped separately in each cell type: a very slow experimental work and, in practice, unmanageable for the whole of an organism.
🧩 The finding: a "fingerprint" common to all cells
The team led by Pulin Li, a researcher at the Whitehead Institute for Biomedical Research and professor of biology at MIT, along with graduate student Nicholas Hutchins, discovered something that changes the picture: each signaling pathway leaves a characteristic fingerprint, a distinctive pattern of genetic activity that reflects the specific signals that the cell received.
The decisive thing is that this fingerprint remains consistent between different cell types for the same signaling pathway. In other words, instead of mapping pathway by pathway and cell type by cell type, scientists can reconstruct the signaling history of many different cells from those shared fingerprints.
|
"Think of speech recognition systems like Siri, which are trained primarily in English, but then use that training to recognize other languages. This is called transfer learning, and it's why IRIS can work on many different cell types." " — Pulin Li, Whitehead Institute / MIT |
⚙️ How IRIS works
IRIS is a neural network-based model: an artificial intelligence system designed to recognize patterns in complex data, similar to how the human brain detects regularities. The program examines a cell's overall gene activity and, from it, estimates which signaling pathways were active—"turned on"—at different points in development.
To train it, Li and Hutchins used an extensive experimental dataset that recorded how thousands of human embryonic stem cells responded to dozens of combinations of six major signaling pathways, at multiple stages of development. The result was a detailed atlas of how signal combinations shape cell behavior.
Then came the litmus test: The researchers applied IRIS to individual cells from mouse embryos during gastrulation, the stage when cells rapidly branch out to very different destinations. The model predicted precisely when and where certain signaling pathways would be activated in cells destined to be part of the heart, gut, muscle, and spinal cord.
|
Target Tissue |
Stage evaluated |
IRIS Result |
|
Heart |
Gastrulation (mouse embryo) |
Accurate prediction of the time and place of signal activation |
|
Intestine |
Gastrulation (mouse embryo) |
Accurate prediction of the time and place of signal activation |
|
Muscle |
Gastrulation (mouse embryo) |
Accurate prediction of the time and place of signal activation |
|
Spinal cord |
Gastrulation (mouse embryo) |
Accurate prediction of the time and place of signal activation |
|
Lung tissue |
Signaling pathway design to induce destiny |
Experimentally Confirmed in Mouse Embryos |
🖼️ View image: Embryonic gastrulation scheme — Wikimedia Commons (Pidalka44, Wikimedia Commons)
🫁 From Prediction to Lung: The Ultimate Test
The most telling experiment came when the team used IRIS to identify which signals would be needed to generate a key cell type in lung development. The model predicted that activating a specific signaling pathway would favor development into lung tissue. The researchers tested that prediction in mouse embryos, and the experimental results confirmed it.
By accurately identifying the signal combinations that drive lung cell development, the team can more reliably generate laboratory models—organoids, miniature three-dimensional structures that mimic real organs—that faithfully reproduce human lung tissue.
|
"In these ways, IRIS is helping us decode the language that cells use to talk to each other, at a much faster speed than we could actually achieve through experiments. " — Nicholas Hutchins, Whitehead Institute |
💊 Why this matters for medicine
The finding, published in the journal Nature Methods, has very concrete implications beyond the developmental biology lab. Once scientists identify the pattern of signals that pushes a stem cell to become a specific cell type, they can artificially recreate those signals to direct the fate of stem cells in laboratory or clinical contexts.
Among the applications that the team highlights are:
• Engineering stem cells for regenerative medicine, guiding their differentiation towards the desired cell type with greater efficiency.
• More precise organoids, that is, three-dimensional models of organs used to study diseases and test drugs before reaching clinical trials.
• Models of lung diseases such as asthma, lung cancer, and pulmonary fibrosis, in which tissue scars often with no known cause.
• Design of regenerative treatments capable, in the future, of repairing damaged lung tissue.
The team notes that these improved models would allow not only a better understanding of why these diseases develop, but also use them as a testing platform for new therapies, with the ultimate goal of designing real regenerative treatments.
🌐 A field in full swing
IRIS joins a recent wave of artificial intelligence models that attempt, each from a different angle, to "read" cell behavior. In early 2026, researchers at Columbia University presented a method capable of accurately predicting gene activity within any human cell from gene expression data from millions of cells in normal tissues, also published in Nature.
Months later, an international consortium made up of the Stowers Institute for Medical Research, Helmholtz Munich, the Technical University of Munich and the University of Oxford presented RegVelo, an AI framework that simultaneously models cell dynamics and gene regulation to predict how cells make decisions about their fate, experimentally validated in zebrafish.
At the same time, other research teams have been developing models such as Sig2Fate, aimed at decoding cell fate from a snapshot of combinatorial signaling in human gastruloids. The common denominator of all this work is the same ambition: to turn traditionally descriptive developmental biology into a science capable of predicting in advance how a cell will behave in the face of a given stimulus.
Specialists in the intersection between artificial intelligence and biology, such as Microsoft researcher Ava Amini, have warned, however, that the path is not without obstacles: several existing models of cellular AI tend to predict only average values rather than actual biological differences, and their performance does not always improve with increasing volume of training data. IRIS, by relying on shared signaling fingerprints between cell types rather than exhaustive mapping, proposes a different way to get around this limitation.
❔ Frequently Asked Questions
❓ What exactly is IRIS?
It is an artificial intelligence model based on neural networks, developed at the Whitehead Institute (affiliated with MIT), which analyzes the gene activity of a cell to estimate which signaling pathways it received and at what point in development.
❓ How does IRIS predict which tissue a cell will become into?
The model detects "fingerprints" of genetic activity specific to each signaling pathway, which remain stable between different cell types. From these fingerprints, it reconstructs the history of signals that the cell received and, with this, predicts its tissue fate: heart, intestine, muscle, spinal cord or lung, among others.
❓ Where and when was the study published?
The results were published in the journal Nature Methods on September 8, 2026, in a study led by Pulin Li and Nicholas Hutchins, from the Whitehead Institute for Biomedical Research.
❓ What is the use of this advance in practice?
It allows for more precise design of protocols to guide stem cells to a desired cell type in the laboratory, improve the manufacture of organoids to study diseases such as asthma, lung cancer or pulmonary fibrosis, and lay the foundations for future regenerative treatments.
❓ Is it the first AI to try to decipher cellular fate?
No. It joins other recent developments, such as a Columbia University model for predicting cell gene activity and RegVelo, a framework developed by institutions such as the Stowers Institute, Helmholtz Munich, and the University of Oxford. What's new about IRIS is its ability to generalize between cell types through transfer learning.
🔮 What's next
The Whitehead Institute team says this approach opens up the possibility of comprehensively mapping the signaling histories of every cell within a mouse or human embryo on a scale that was previously unattainable. If that mapping comes to fruition, the promise is twofold: to understand in unprecedented detail how a handful of undifferentiated cells build an entire body, and to learn how to guide that process from the lab to repair what the disease breaks.